The Emergy of a Prompt

In the AI Story

The typical user's experience of AI — fast, responsive, essentially costless at the margin — is the product of deliberate interface design. Every layer of abstraction, from the chat window to the API, was engineered to hide the substrate. The prompt travels through fiber optic cables manufactured from ultrapure glass; the response is generated by GPUs fabricated in facilities costing twenty billion dollars each; the electricity is drawn from grids sustained by fossil fuels, nuclear fission, or renewables whose own emergy chains extend across decades of infrastructure investment.

Lawrence Berkeley National Laboratory projects U.S. data center electricity consumption will grow from 176 terawatt-hours in 2023 to between 325 and 580 terawatt-hours by 2028 — a doubling or tripling driven almost entirely by AI workloads. A single AI-focused facility can consume millions of gallons of water daily for cooling, drawn from aquifers that recharge over centuries. These are not externalities in the economist's sense. They are the actual structure of the transaction that the interface presents as weightless.

The deepest layer of the emergy chain is training data. Every text in the corpus represents the endpoint of an investment extending through the author's education, the institutions that supported that education, the agricultural surplus that freed her from subsistence, the printing infrastructure that distributed her work, and the theoretical frameworks developed by thousands of prior researchers, each with their own emergy chain. When a model generates a response synthesizing such sources, it performs an emergy drawdown on centuries of accumulated intellectual capital.

Odum's method refuses the comforting distinction between primary cost and invisible substrate. The amplifier metaphor in Edo Segal's Orange Pill captures the experience but not the physics. The imagination-to-artifact ratio has not collapsed; it has been subsidized. The subsidy comes from geological reserves, technological infrastructure, and civilizational intellectual capital. Seeing this does not forbid the tools. It forbids using them as though the substrate were infinite.

Origin

Odum developed the emergy methodology in the 1970s and spent four decades demonstrating that systems which appear efficient often operate through deep, unaccounted subsidies. His work on industrial agriculture produced the canonical example: between five and fifteen calories of fossil fuel energy consumed for every calorie of food delivered to a human mouth.

The application to AI — sketched in this volume — extends Odum's framework into territory he anticipated but did not live to see. In 1973, he placed 'computer and human information processing' at the apex of the energy hierarchy, identifying where the twenty-first century would find itself decades before the infrastructure existed to put any system there.

Key Ideas

Five-layer chain. Minerals, infrastructure, electricity, water, training data — each a transformation with its own transformity, none visible at the interface.

Water consumption is consequential. Data center cooling draws from aquifers whose recharge timescales are orders of magnitude longer than the consumption timescales.

Interface design conceals substrate. The experience of frictionless creation is itself an engineered output, not a description of the transaction.

Training data is the deepest layer. Centuries of civilizational intellectual labor constitute an input unlike any other: nonrenewable at the timescale of use, irreplaceable through technical means.

Accounting does not forbid use. Odum was not a Luddite. The framework demands that systems know their costs so they can be managed, not that they be abandoned.

Debates & Critiques

Industry responses typically argue that efficiency gains will reduce the per-prompt cost over time, and that renewables will displace fossil fuels in the grid mix. Both claims are partially true and partially beside the point: emergy accounting measures total throughput, not efficiency per unit, and the maximum power principle predicts that systems will grow demand at least as fast as efficiency gains permit.